infinite values PSI

[gverbock]

Problem Description
In the current implementation of AutoDist, the PSI is infinite if one of the bin from the test dataset is empty. Is could be nice to include the possibility to as add one 'fake' observation to the empty bin and get an estimate of the PSI that is closer to reality

Desired Outcome

Solution Outline

class DistributionStatistics(object):


def compute(self, d1, d2, avoid_empty_bins, verbose=False)

if self.binning_strategy:
            self.binner.fit(d1)
            d1_preprocessed = self.binner.counts_
            d2_preprocessed = self.binner.compute(d2)
            if avoid_empty_bins: 
                        d2_preprocessed = np.where(d2_preprocessed == 0, 1, d2_preprocessed)

Another possibility is to adjust the PSI function itself:

def psi(d1, d2, verbose=False):



# Eventually correct for empty bins in d2
     if avoid_empty_bins: 
                        previous_count = d2.sum()
                        d2 = np.where(d2 == 0, 1, d2)

# Calculate the number of samples in each distribution
    n = d1.sum()
    m = d2.sum()
    if previous_count != m:
           # log warning

 # Calculate the ratio of samples in each bin
    expected_ratio = d1 / n
    actual_ratio = d2 / m

 psi_value = np.sum((actual_ratio - expected_ratio) * np.log(actual_ratio / expected_ratio))

Also it could be good to have a warning when avoid_empty_bin is effectively used.

In this dummy example it leads to a PSI of zero but in larger dataset, the impact is limited.